Pattern Recognition 40 (2007) 3146 – 3151 www.elsevier.com/locate/pr A hybrid wavelet-based ﬁngerprint matcher Loris Nanni ∗ , Alessandra Lumini DEIS, IEIIT–CNR, Università di Bologna, Viale Risorgimento 2, 40136 Bologna, Italy Received 19 June 2006; received in revised form 3 January 2007; accepted 27 February 2007 Abstract In this work we present a hybrid ﬁngerprint matcher system based on the multi-resolution analysis of the ﬁngerprint pattern and on minutiae- based registration module. Two ﬁngerprints are ﬁrst aligned using their minutiae, then the images are divided in sub-windows and each sub- window is decomposed into frequency sub-bands at different decomposition levels using a set of wavelet functions, ﬁnally a distinct classiﬁer is trained on each sub-band to distinguish matching pairs of ﬁngerprint from non-matching one (deﬁning a two-class matching problem). The features extracted for the matching are the standard deviation of the image convolved with 16 Gabor ﬁlters. The selection among the pool of matchers, is performed by running Sequential Forward Floating Selection. The retained matchers are weighted by a novel localized quality measure and combined by a fusion rule. Extensive experiments conducted over the four FVC2002 ﬁngerprint databases show the effectiveness of the proposed approach.  2007 Pattern Recognition Society. Published by Elsevier Ltd. All rights reserved. Keywords: Fingerprint matching; Hybrid approach; Image based; Minutiae registration 1. Introduction Fingerprint automatic veriﬁcation have been widely stud- ied in the literature and the various approaches proposed may be broadly classiﬁed as minutiae-based, correlation-based or image-based (for a good survey see Ref. [1]). Minutiae-based approaches ﬁrst extract the minutiae from the ﬁngerprint im- ages; then, the matching between two ﬁngerprints is made us- ing the two sets of minutiae locations. The performance of minutiae-based techniques rely on the accuracy of the minu- tiae detection process and the use of sophisticated aligning and matching techniques to compare two minutiae sets. In correlation-based ﬁngerprint matching [2], the template and query ﬁngerprint images are spatially correlated to estimate the degree of similarity between them. Image-based approaches extract a set of numerical features directly from the grey-level image of the ﬁngerprint and the matching decision among two ﬁngerprints is made using only these features. Image-based approaches may be the only viable choice in particularly difﬁcult cases, when the quality of the ﬁngerprint is too low to allow reliable minutia extraction. ∗ Corresponding author. Tel.: +393493511673. E-mail address: lnanni@deis.unibo.it (L. Nanni). 0031-3203/$30.00  2007 Pattern Recognition Society. Published by Elsevier Ltd. All rights reserved. doi:10.1016/j.patcog.2007.02.018 A well known image-based technique is proposed in Ref. [3], where a compact, ﬁxed length, feature vector, named Finger- Code, is adopted to capture global and local features of a ﬁngerprint. This technique makes use of the texture features available in a ﬁngerprint to compute the feature vector by ap- plying Gabor ﬁlters in the region around the ﬁngerprint core point. In Ref. [4] a FingerCode variant has been proposed where the ﬁngerprint images were aligned using the overall minu- tiae information; this approach is more robust than using only the core point for aligning image pairs. In Ref. [5] a method based on an integrated Wavelet and Fourier–Mellin transformed feature (WFMT) is proposed, where the only pre-processing step is the reference point detection in the ﬁngerprint image; this approach improves the previous ones thanks to use of both the wavelet transform, which preserves the local edges and grants a noise reduction in the low-frequency domain, and of the Fourier–Mellin transform, which produces a translation, rotation and scale invariant feature vector. Anyway, the performance of almost all image-based tech- niques is affected by non-linear distortions and noise present in the image. Several techniques have been suggested in the literature to handle such distortions (see Refs. [1, Section 4.5, 6–8]). In our previous works [6–8], the region of interest around the reference point (the core point) was partitioned in smaller